302 lines
10 KiB
Python
302 lines
10 KiB
Python
import re
|
||
import pandas as pd
|
||
import math
|
||
from functools import reduce
|
||
import argparse
|
||
import os
|
||
import sqlite3
|
||
from itertools import chain
|
||
|
||
model_filename = "model.db"
|
||
|
||
def genmod():
|
||
corpus_path = "./corpus/"
|
||
df_list = []
|
||
for file in os.listdir(corpus_path):
|
||
if file.endswith(".csv"):
|
||
df = pd.read_csv(corpus_path+file, header=0, names=['hanji', 'lomaji'])
|
||
df_list.append(df)
|
||
df = pd.concat(df_list)
|
||
df['lomaji'] = df['lomaji'].str.lower()
|
||
|
||
new_data = []
|
||
|
||
for index, row in df.iterrows():
|
||
hanji = list(filter(lambda x : re.match("[^、();:,。!?「」『』]", x), list(row['hanji'])))
|
||
tl = re.split(r'(:?[!?;,.\"\'\(\):]|[-]+|\s+)', row['lomaji'])
|
||
tl2 = list(filter(lambda x : re.match(r"([^\(\)^!:?; \'\",.\-\u3000])", x), tl))
|
||
new_data.append((hanji, tl2))
|
||
if (len(hanji) != len(tl2)):
|
||
raise ValueError(f"length of hanji {hanji} is different from romaji {tl2}.")
|
||
|
||
#model_filename = "model.db"
|
||
try:
|
||
os.remove(model_filename)
|
||
except OSError:
|
||
pass
|
||
|
||
con = sqlite3.connect(model_filename)
|
||
cur = con.cursor()
|
||
cur.execute("CREATE TABLE pronounce(hanji, lomaji, freq)")
|
||
|
||
|
||
char_to_pronounce = {}
|
||
|
||
for i in new_data:
|
||
hanji = i[0]
|
||
lomaji = i[1]
|
||
for j in range(len(i[0])):
|
||
if not hanji[j] in char_to_pronounce:
|
||
char_to_pronounce[hanji[j]] = {lomaji[j] : 1}
|
||
elif not lomaji[j] in char_to_pronounce[hanji[j]]:
|
||
char_to_pronounce[hanji[j]][lomaji[j]] = 1
|
||
else:
|
||
char_to_pronounce[hanji[j]][lomaji[j]] += 1
|
||
|
||
|
||
for i in char_to_pronounce.keys():
|
||
hanji = char_to_pronounce[i]
|
||
for j in hanji.keys():
|
||
cur.execute("INSERT INTO pronounce VALUES(?, ?, ?)", (i,j, hanji[j]))
|
||
|
||
all_chars = char_to_pronounce.keys()
|
||
init_freq = {} #詞kap句開始ê字出現次數
|
||
cur.execute("CREATE TABLE initial(char, freq)")
|
||
|
||
|
||
for i in new_data:
|
||
head_hanji = i[0][0]
|
||
|
||
if head_hanji in init_freq:
|
||
init_freq[head_hanji] += 1
|
||
else:
|
||
init_freq[head_hanji] = 1
|
||
|
||
#補字
|
||
min_weight = 0.1
|
||
|
||
for i in all_chars:
|
||
if not i in init_freq.keys():
|
||
init_freq[i] = 0.1
|
||
|
||
for i in init_freq.keys():
|
||
cur.execute("INSERT INTO initial VALUES(?, ?)", (i, init_freq[i]))
|
||
|
||
char_transition = {}
|
||
cur.execute("CREATE TABLE transition(prev_char, next_char, freq)")
|
||
|
||
for i in new_data:
|
||
hanji = i[0]
|
||
for j in range(len(i[0])-1):
|
||
this_hanji = hanji[j]
|
||
next_hanji = hanji[j+1]
|
||
if not this_hanji in char_transition:
|
||
char_transition[this_hanji] = {next_hanji : 1}
|
||
elif not next_hanji in char_transition[this_hanji]:
|
||
char_transition[this_hanji][next_hanji] = 1
|
||
else:
|
||
char_transition[this_hanji][next_hanji] += 1
|
||
|
||
for i in char_transition.keys():
|
||
next_char = char_transition[i]
|
||
for j in next_char.keys():
|
||
cur.execute("INSERT INTO transition VALUES(?, ?, ?)", (i, j, next_char[j]))
|
||
|
||
|
||
#get_homophones("lí", cur, con)
|
||
|
||
con.commit()
|
||
con.close()
|
||
|
||
def get_homophones(pron, cur, con):
|
||
homophones_raw = cur.execute("select hanji FROM pronounce where lomaji = ?", (pron, )).fetchall()
|
||
homophones = list(map(lambda x: x[0], homophones_raw))
|
||
|
||
return homophones
|
||
|
||
def convert(sentences):
|
||
splitted = re.split(r'(:?[!?;,.\"\'\(\):])', sentences)
|
||
splitted_cleaned = list(filter(lambda x : x != '', splitted))
|
||
|
||
result = list(map(lambda s : convert_one_sentence(s), splitted_cleaned))
|
||
|
||
flatten_result = [x for xs in result for xss in xs for x in xss]
|
||
result_string = "".join(flatten_result)
|
||
|
||
|
||
print(result_string)
|
||
return result_string
|
||
|
||
def convert_one_sentence(sentence):
|
||
full_width = ["!", "?", ";",":",",","。", "(", ")"]
|
||
half_width = ["!", "?", ";", ":", ",", ".", "(", ")"]
|
||
|
||
if len(sentence) == 1:
|
||
for i in range(len(half_width)):
|
||
if sentence[0] == half_width[i]:
|
||
return [[full_width[i]]]
|
||
|
||
|
||
weight = 2/3
|
||
|
||
splitted = re.split(r'(--?|\s+)', sentence)
|
||
filtered = list(filter(lambda x :not re.match(r'(--?|\s+)', x), splitted))
|
||
small_capized = list(map(lambda x : x.lower(), filtered))
|
||
print("======", small_capized)
|
||
con = sqlite3.connect(model_filename)
|
||
cur = con.cursor()
|
||
|
||
homophones_sequence_raw = list(map(lambda x : get_homophones(x, con, cur), small_capized))
|
||
|
||
homophones_sequence = [list(map (lambda x : {"char": x,
|
||
"prev_char": None,
|
||
"prob" : 1}, i)) for i in homophones_sequence_raw]
|
||
|
||
|
||
|
||
head_freqs = list(map(lambda x : x[0], cur.execute('''select initial.freq FROM initial
|
||
INNER JOIN pronounce ON pronounce.hanji = initial.char
|
||
WHERE pronounce.lomaji = ?''', (small_capized[0], )).fetchall()))
|
||
|
||
return_result = [None] * len(small_capized)
|
||
|
||
if head_freqs == []:
|
||
return_result[0] = filtered[0]
|
||
homophones_sequence[0] = [{"char": filtered[0],
|
||
"prev_char": None,
|
||
"prob" : 1}]
|
||
|
||
else:
|
||
head_freq_total = reduce(lambda x , y : x + y, head_freqs)
|
||
|
||
for i in homophones_sequence[0]:
|
||
i_freq = cur.execute('''select initial.freq FROM initial
|
||
WHERE initial.char = ?''', (i['char'])).fetchall()[0][0]
|
||
|
||
i['prob'] = i_freq / head_freq_total
|
||
print(i)
|
||
|
||
#for i in homophones_sequence[0]:
|
||
|
||
print("+++++", return_result)
|
||
|
||
if len(small_capized) == 1:
|
||
max_prob = -math.inf
|
||
max_prob_char = None
|
||
for i in homophones_sequence[0]:
|
||
if i['prob'] > max_prob:
|
||
max_prob_char = i['char']
|
||
max_prob = i['prob']
|
||
|
||
return_result[0] = max_prob_char
|
||
|
||
else:
|
||
for i in range(1,len(small_capized)):
|
||
char_freqs = list(map(lambda x : x[0], cur.execute('''select initial.freq FROM initial
|
||
INNER JOIN pronounce ON pronounce.hanji = initial.char
|
||
WHERE pronounce.lomaji = ?''', (small_capized[i], )).fetchall()))
|
||
|
||
if char_freqs == []:
|
||
return_result[i] = filtered[i]
|
||
homophones_sequence[i] = [{"char": filtered[i],
|
||
"prev_char": None,
|
||
"prob" : 1}]
|
||
prev_char = ""
|
||
max_prob = -math.inf
|
||
for m in homophones_sequence[i-1]:
|
||
if m['prob'] > max_prob:
|
||
max_prob = m['prob']
|
||
prev_char = m['char']
|
||
homophones_sequence[i][0]['prob'] = max_prob
|
||
homophones_sequence[i][0]['prev_char'] = prev_char
|
||
else:
|
||
total_transition_freq = cur.execute('''
|
||
SELECT sum(t.freq)
|
||
FROM transition as t
|
||
INNER JOIN pronounce as p1 ON p1.hanji = t.prev_char
|
||
INNER JOIN pronounce as p2 ON p2.hanji = t.next_char
|
||
where p2.lomaji = ? and p1.lomaji = ?''',
|
||
(small_capized[i], small_capized[i-1])).fetchall()[0][0]
|
||
for j in homophones_sequence[i]:
|
||
prev_char = None
|
||
max_prob = -math.inf
|
||
|
||
for k in homophones_sequence[i-1]:
|
||
k_to_j_freq_raw = cur.execute('''select freq from transition
|
||
where prev_char = ? and next_char = ? ''', (k["char"], j["char"])).fetchall()
|
||
if k_to_j_freq_raw == []:
|
||
den = cur.execute('''
|
||
SELECT sum(p.freq)
|
||
FROM pronounce as p
|
||
inner join pronounce as p2
|
||
on p.hanji = p2.hanji where p2.lomaji = ?''', (small_capized[i],)).fetchall()[0][0]#分母
|
||
#分子
|
||
num = cur.execute(''' SELECT sum(freq) FROM pronounce as p where hanji = ?''', (j["char"],)).fetchall()[0][0]
|
||
print("+++", num, den)
|
||
k_to_j_freq = num/den * (1-weight)
|
||
|
||
else:
|
||
num = k_to_j_freq_raw[0][0]
|
||
don = total_transition_freq
|
||
k_to_j_freq =num/don * weight
|
||
print("k_to_j_fr", k["char"], j["char"], k_to_j_freq)
|
||
if k_to_j_freq * k["prob"] > max_prob:
|
||
max_prob = k_to_j_freq * k["prob"]
|
||
prev_char = k["char"]
|
||
print("~-~_~-~-~-~-", prev_char, j["char"], max_prob)
|
||
j["prob"] = max_prob
|
||
j["prev_char"] = prev_char
|
||
|
||
max_prob = -math.inf
|
||
current = ""
|
||
prev_char = ""
|
||
for i in homophones_sequence[len(homophones_sequence)-1]:
|
||
if i["prob"] > max_prob:
|
||
max_prob = i["prob"]
|
||
current = i["char"]
|
||
prev_char = i["prev_char"]
|
||
|
||
print("~tail~~", current)
|
||
print(homophones_sequence)
|
||
return_result[len(homophones_sequence)-1] = current
|
||
|
||
for i in range(len(homophones_sequence)-2, -1, -1):
|
||
current_ls = list(filter(lambda x : x["char"] == prev_char,
|
||
homophones_sequence[i]))
|
||
print(prev_char)
|
||
return_result[i] = prev_char
|
||
current = current_ls[0]["char"]
|
||
prev_char = current_ls[0]["prev_char"]
|
||
|
||
|
||
print(return_result)
|
||
|
||
return return_result
|
||
|
||
|
||
def poj_to_tl(sentence):
|
||
return sentence
|
||
|
||
parser = argparse.ArgumentParser()
|
||
parser.add_argument('--genmod', help='generate the model', action='store_true',
|
||
required=False,)
|
||
|
||
parser.add_argument('sentence', metavar='SENTENCE', nargs='?',
|
||
help='the sentence to be converted')
|
||
parser.add_argument('--form', metavar='FORM', choices=["poj", "tl"], nargs=1,
|
||
default=['poj'],
|
||
help='the orthography to be used (poj or tl). Default is poj.')
|
||
|
||
args = parser.parse_args()
|
||
|
||
if args.genmod == True:
|
||
genmod()
|
||
elif args.sentence != None:
|
||
if args.form == ['poj']:
|
||
sentence = poj_to_tl(args.sentence)
|
||
convert(sentence)
|
||
else:
|
||
convert(args.sentence)
|
||
else:
|
||
parser.print_help()
|
||
|